PaperlayerView exact revision on arXiv

Source-linked AI summary

Cost-Driven Offloading for DNN-based Applications over Cloud, Edge and End Devices

Bin Lin, Yinhao Huang, Jianshan Zhang, Junqin Hu, Xing Chen, Jun Li

arXiv:1907.13306v1cs.DCeess.SP

TL;DR

DNNs deployed in the cloud can suffer substantial input-transfer delay, creating a cost-aware offloading problem across cloud, edge, and end devices under deadline constraints. The paper uses PSO-GA with DNN preprocessing and GA operators to optimize this offloading, and reports lower system cost than benchmark strategies.

  • Problem

    Cloud deployment can cause serious input-transfer delay, while hybrid offloading must minimize transmission and layer-execution cost under DNN deadlines.

  • Method

    The paper combines DNN layer preprocessing with self-adaptive discrete PSO and GA crossover and mutation operators for cost-driven hybrid offloading.

  • Results

    PSO-GA effectively reduces offloading system cost relative to benchmark strategies while completing DNN layers within their corresponding deadlines.

  • Takeaways & Limitations

    Cost-driven PSO-GA provides a supported strategy for deadline-constrained DNN layer offloading across cloud, edge, and end devices.

Abstract

from arXiv · show

Currently, deep neural networks (DNNs) have achieved a great success in various applications. Traditional deployment for DNNs in the cloud may incur a prohibitively serious delay in transferring input data from the end devices to the cloud. To address this problem, the hybrid computing environments, consisting of the cloud, edge and end devices, are adopted to offload DNN layers by combining the larger layers (more amount of data) in the cloud and the smaller layers (less amount of data) at the edge and end devices. A key issue in hybrid computing environments is how to minimize the system cost while accomplishing the offloaded layers with their deadline constraints. In this paper, a self-adaptive discrete particle swarm optimization (PSO) algorithm using the genetic algorithm (GA) operators was proposed to reduce the system cost caused by data transmission and layer execution. This approach considers the characteristics of DNNs partitioning and layers offloading over the cloud, edge and end devices. The mutation operator and crossover operator of GA were adopted to avert the premature convergence of PSO, which distinctly reduces the system cost through enhanced population diversity of PSO. The proposed offloading strategy is compared with benchmark solutions, and the results show that our strategy can effectively reduce the cost of offloading for DNN-based applications over the cloud, edge and end devices relative to the benchmarks.

I. INTRODUCTION

DNN deployment in the cloud can impose serious input-transfer delay for resource-constrained IoT devices, motivating hybrid cloud-edge-end-device offloading. The paper proposes PSO-GA to minimize transmission and layer-computing costs while meeting application deadlines.

  • Cloud deployment causes serious input-transfer delay, while IoT devices lack the energy and computing resources for computation-intensive DNNs.
  • Hybrid environments partition DNNs across cloud, edge, and end devices, placing larger layers in the cloud and smaller layers nearer the devices.This arrangement is described as reducing core-network traffic and transmission delay while improving system accuracy.
  • The unresolved problem is minimizing transmission and computation cost while scheduling DNN layers within application deadlines across heterogeneous cloud and edge servers.
  • PSO-GA combines self-adaptive discrete PSO with GA operators to reduce offloading cost while respecting all DNN application deadlines.
  • Preprocessing merges eligible adjacent layers, reducing the number of PSO particle dimensions before optimization.

II. RELATED WORK

Prior work largely targets response-delay reduction through edge or hybrid DNN deployment, while system cost from both data transmission and layer execution remains insufficiently addressed. This paper positions cost-aware deadline-constrained offloading in hybrid environments as the open issue.

  • Existing DNN offloading studies mostly reduce response delay by sending layers from resource-constrained devices to edge servers.
  • Hybrid cloud-edge-end-device approaches partition DNNs or distribute them across computing hierarchies, but prior studies give limited attention to layer-offloading decisions.
  • Prior work considers communication cost in some distributed-DNN settings but may ignore layer-computing cost, while workflow methods do not account for MEC.
  • The open issue is minimizing data-transmission and layer-execution cost while meeting DNN deadlines in hybrid computing environments.

III. PROBLEM DEFINITION AND ANALYSIS

The paper defines its objective as minimizing the cost of transmitting data and executing DNN layers while satisfying each DNN application's deadline constraint.

  • The objective is to minimize data-transmission and layer-execution cost during DNN offloading while satisfying each application's deadline.

A. Problem Definition

The problem defines DNN offloading across cloud, edge, and end-device servers as minimizing transmission and execution cost while meeting application deadlines. It models DNN dependencies, server capabilities, data movement, and execution constraints for this optimization.

  • Computing environment: The hybrid environment contains cloud servers with strong computing power, edge servers with general power, and end devices with poor power.Server computing power and cost are modeled per second, with computing power assumed known and stable.
  • Data transmission: Transmission time and cost depend on dataset size, bandwidth, and the source and destination servers.Bandwidth is assumed known and non-fluctuating; end devices communicate with nearby edge servers through WiFi rather than directly with one another.
  • DNN model: Each DNN is a directed acyclic graph of layers, dependencies, and datasets, with feasibility defined by completing within its deadline.Datasets include inputs, intermediate outputs, and final outputs, while layer execution follows dependency order.
  • Execution model: Execution uses a serial server model in which one server runs one layer at a time and each layer remains on a single server.A layer may require multiple datasets from different servers, and datasets may be shared by multiple layers.
  • Objective: The offloading strategy minimizes total data-transmission and layer-execution cost while ensuring every DNN meets its deadline.A layer executes only after it is assigned to a server and all required datasets reach that server.

IV. OFFLOADING STRATEGY BASED ON PSO-GA

The PSO-GA offloading strategy searches for a minimum-cost layer-to-server mapping that satisfies DNN deadlines. It preprocesses DNNs by merging eligible adjacent layers to reduce optimization dimensionality, although this can affect the final offloading result.

  • Optimization objective: PSO-GA finds a mapping of all DNN layers to hybrid-computing servers that minimizes total cost under completion-time deadlines.The optimization targets the system cost while requiring each DNN completion time to remain within its corresponding deadline.
  • Preprocessing: Preprocessing merges adjacent layers when the precursor out-degree and successor in-degree are both 1, removing their common cut-edge.The operation compresses layer and data-transmission counts before PSO-GA optimization.
  • Preprocessing: GoogleNet’s compressed layer count reaches about 48%, and fewer layers improve PSO-GA efficiency by reducing particle encoding dimensionality.The particle dimension depends on the total number of layers.
  • Preprocessing trade-off: Preprocessing may change the final offloading result because merged layers must be assigned to the same server.For example, merging FC1 and FC2 prevents assigning those layers to two different servers.

B. PSO-GA

PSO-GA encodes DNN offloading as discrete particles and combines PSO updates with GA crossover, mutation, preprocessing, and adaptive search to avoid premature convergence while respecting deadlines.

  • Problem Encoding: PSO-GA represents each DNN offloading strategy as a particle encoding server assignments and processing orders for all layers.The encoding uses server-order nesting; each particle has 2p dimensions for p layers and distinguishes feasible from infeasible deadline satisfaction.
  • Fitness Function: The fitness function prioritizes feasible particles by lower system cost, while comparing infeasible particles by total completion time.When one particle is feasible, it is selected over an infeasible particle; when both are infeasible, the lower total completion time is preferred.
  • Update Strategy: GA crossover and mutation update PSO cognition and inertia components to refresh particles and mitigate premature convergence.Crossover replaces a particle segment with the corresponding segment from pBest or gBest, while mutation changes a server assignment at a randomly selected location.
  • Adaptive Search: The adaptive mechanism adjusts search ability according to the difference between the global-best and current particles, addressing nonlinear DNN offloading behavior.The paper states that the conventional linear inertia adjustment does not match the nonlinear characteristics of DNN layer offloading.

V. EXPERIMENTAL RESULTS AND ANALYSIS

The experiments use fixed PSO-GA parameters and report server configurations and costs through a dedicated table for the evaluation environment.

  • Parameter Settings: The evaluation sets PSO-GA population size to 100, maximum iterations to 1000, and inertia and acceleration parameters to specified start and end values.The reported settings are wmax=0.9, wmin=0.4, c1 from 0.9 to 0.2, and c2 from 0.4 to 0.9.
  • Server Configuration: Table IV presents the configurations and costs of all servers used in the hybrid computing environment.The supplied passage identifies the table scope but does not provide its cell values.

A. Experimental Setup

The evaluation covers four DNN architectures in a 20-server cloud-edge-end environment, using multiple deadline settings derived from HEFT execution times.

  • DNN Workloads: Experiments evaluate AlexNet, VGG19, GoogleNet, and ResNet101, whose structures, datasets, execution times, and inter-layer transmission amounts differ.The tested DNN data are recorded in a file referenced by the paper.
  • Computing Environment: The hybrid environment contains 20 servers divided among 10 end devices, 5 edge servers, and 5 cloud servers.End servers have the lowest configurations and execute DNN layers without charging; server processing capacity within a category is roughly proportional to cost.
  • Deadline Settings: Each DNN is tested under five deadlines set as 1.2, 1.5, 3, 5, or 8 times its HEFT execution time.The deadline multiplier is applied to H(G_i), the DNN execution time based on HEFT.

B. Competitive Algorithms

PSO-GA is compared with modified GA, Greedy, and prePSO algorithms adapted to cost-driven DNN offloading in hybrid computing environments.

  • Compared Algorithms: The evaluation compares PSO-GA with modified GA, Greedy, and prePSO algorithms.GA and Greedy are adapted for hybrid-environment cost-driven offloading, while prePSO is the PSO-GA variant with preprocessing discussed earlier.
  • Genetic Algorithm: GA evaluates offloading chromosomes using computation and transmission costs for each DNN layer.Its binary encoding uses the same feasibility-oriented fitness cases defined by the paper.
  • Greedy Algorithm: Greedy assigns each layer to the cheapest server that can satisfy the deadline, escalating to the second-cheapest server when necessary.The procedure iterates this selection across layers.

C. Experimental Results and Analysis

Across single- and multi-DNN experiments, PSO-GA generally achieves the strongest cost performance, while looser deadlines and greater edge computing power reduce system cost and improve feasibility.

  • PSO-GA has the best overall performance across the evaluated offloading strategies, outperforming alternatives as deadlines are relaxed.For one DNN per device, it is attributed to iterative global evolution; for three DNNs per device, it also maintains the best performance.
  • System cost decreases as deadline constraints become looser because more layers can be assigned to cheaper servers.
  • In the one-DNN experiments, AlexNet has lower system costs than VGG19 and ResNet101, while GoogleNet has no feasible strategy before D3(G).
  • With three DNNs per end device, system cost is almost 4 times that of the one-DNN-per-device setting.The three-DNN experiments also use deadlines twice as long to reduce infeasible solutions.
  • For three AlexNets per device, Greedy becomes extremely unsuitable as layer count increases, whereas PSO-GA has the best performance.
  • Increasing edge-server computing power improves system cost 4% to 31% more than increasing cloud-server computing power.AlexNet places most layers on edge servers, and its layer computing cost accounts for most total system cost in this experiment.

D. Industrial Applications

The industrial application targets UAV-swarm package delivery, where DNN-based computer vision must support real-time decisions under limited battery capacity and latency constraints.

  • UAV-swarm package delivery uses hybrid cloud-edge-end computing to place larger, high-business-intelligence layers in the cloud and smaller layers nearer the devices.

VI. CONCLUSION

The paper concludes that PSO-GA reduces deadline-constrained DNN offloading cost across cloud, edge, and end devices, while identifying bandwidth variation and server-specific price/performance ratios as future considerations.

  • PSO-GA effectively reduces offloading system cost within each DNN’s deadline across cloud, edge, and end devices.
  • Looser deadlines allow more layers to use cheaper servers, and sufficiently loose deadlines permit all layers to remain on their original end devices without data transmission.
  • The study does not yet account for bandwidth changes between servers in different regions or different server-specific price/performance ratios.
Loading 1907.13306v1…